Deep learning solution from low resolution multi-view images

I have an image grid of multiple images as seen in the attached image. The image is rendered by using a smaller camera that shifts. The images are low resolution 100x100 or smaller and the baselines between them are narrow. I have used semi-global block-matching and wls filter to compute the disparity after I up-sample the images.

I would like to improve my algorithm by using a deep-learning solution using a binocular disparity that would improve the result, especially for more complex images.

Please only contact me if you work on a similar topic before.